Jensen Huang is betting $250 billion on a hallucination. The hallucination is that the large language model—and it is worth being precise about what “the algorithm” actually is, because the public discourse has the misleading habit of treating it as a thing rather than a continually-tuned set of weights serving a continually-revised objective function—will become a general-purpose reasoning engine that generates revenue at a multiple of its planetary-scale capital cost. The check, reported on Monday by the Wall Street Journal, is a backstop for OpenAI, which is attempting to lease a ten-gigawatt data-center project in southern Ohio being built by SoftBank’s energy subsidiary at a total price that could exceed half a trillion dollars. That is more than the annual nominal GDP of Norway. It is being spent on a technology whose leading commercial application, to date, is summarizing emails that nobody reads.

The transaction is a specimen of what the novelist and essayist Cory Doctorow calls the bezzle: the interval, borrowed from Galbraith, between the commission of a financial fraud and its discovery, during which the perpetrator has the gain and the victim feels no loss. The bezzle, in the AI bubble, is the period stretching from the moment Nvidia signs the backstop to the moment it becomes inescapably obvious that the revenue projections that justify a $250 billion guarantee were constructed in a spreadsheet whose assumptions would not survive an audit by a competent millwright. Main Street Independent has already done the arithmetic: as a base-rate analysis put it in June, the revenue multiple being priced into current AI investments is historically unprecedented—far beyond the capital-formation patterns that preceded the great tech busts of the early 2000s and the late 1990s. As the newsletter has also detailed, the FOMO driving Wall Street pros into this rally makes professional skeptics scarce, letting the bezzle swell with little friction. What is new is not the hype, which is perennial, but the scale, which is now driven by a handful of firms that control the physical layer—the chips, the data-center power purchase agreements, the submarine cables—of the internet’s next generation.

The same Wall Street Journal newsletter that disclosed the Nvidia backstop also noted that predictions of an AI-driven hiring wipeout have yet to materialize. Companies across tech, transportation, and defense are hiring again. That is the “truck freight volumes are up” argument for the bubble not being a bubble. It is, to be precise about it, a lagging indicator. The hiring that is being celebrated is, overwhelmingly, the hiring of engineers to build the data centers and lay the fiber that the AI capex wave demands. It is the Canadian manufacturing-employment bump in the mid-2000s, when the steel mill was running three shifts to fill pipeline orders that would evaporate before the Christmas shutdown. My father, a journeyman millwright at Manitoba Rolling Mills, spent thirty years watching the same pattern: a capital-expenditure boom that felt like prosperity until the orders stopped and the mill went to a single skeleton shift for a decade. The thing being built was real; the assumption that the thing being built would be paid for by a commensurate flow of future income was not.

The Federal Reserve, which will deliver an “unusually unpredictable” interest-rate decision on the same day Microsoft and Meta report earnings, is the institutional expression of that distinction. The central bank’s job is to assess whether the economy can absorb the price level the financial system has already priced in. If the Fed judges that the AI capex wave is adding demand that will outstrip supply, it will keep rates higher for longer, tightening the noose around the discount rate used to value future AI revenues. If it judges that the wave is, as the tech-industry lobby insists, productivity-enhancing supply-side wizardry that will pay for itself through deflationary magic, it will cut rates and extend the bezzle. The Fed’s staff, who are not paid to take sides in theological disputes about the imminence of artificial general intelligence, will look at the same chart of cumulative AI-industry capital spending and cumulative AI-industry revenue that everyone else is looking at. The chart, to an observer trained in formal protocol verification, looks like a proof sketch in which the critical lemma—revenue catches up to capex before the bond covenants mature—is asserted without demonstration. The lemma is the hallucination. The rest of the proof is sound engineering, and it does not matter.

Microsoft and Meta are expected to report AI-related revenue streams that have yet to scale proportionally to their capital expenditures, making these earnings reports the latest test of whether the revenue gap is closing. The Fed’s decision is unusually unpredictable because core PCE remains sticky while credit spreads have tightened to levels that otherwise signal monetary easing—a conflict that makes the central bank’s signal especially hazardous for a market already priced for a pivot.

There is a move, in the tech-policy debate, to treat the AI bubble as a morality tale about gullible investors. That framing is comfortable because it moralizes a structural failure. The failure is not that investors are credulous; the failure is that the same four platform firms that control the cloud layer, the mobile-operating-system duopoly, and the international submarine-cable capacity—Google, Microsoft, Amazon, and Meta—are also the primary customers for Nvidia’s chips and the primary lessors of the data-center capacity that SoftBank is building. The money flowing into the Ohio project is not being wagered on an unknown startup; it is being routed, through intricate contractual cascades, back to the balance sheets of firms that already dominate the internet’s physical infrastructure. The bezzle, in this configuration, is not money lost to the wind; it is money transferred from the equity portfolios of index-fund holders into the capex lines of hyperscaler income statements, and from there into the revenue line of Nvidia, which is, at this writing, the most valuable semiconductor company in the history of the industry. The hallucination is paying Jensen Huang’s bills. The question is whether it will pay the index-fund holders whose retirement accounts are being used to underwrite it.

The Wall Street Journal piece mentions that more than one hundred House Democrats voted to cut off military financing for Israel, that children are leaving large cities, that Waymo’s robotaxis have racked up parking tickets in Austin. These are not digressions. They are the ordinary noise of a country that is simultaneously funding a half-trillion-dollar data center on the assumption that a next-word predictor will become a conscious being, cutting children’s community institutions out of its largest cities, and billing autonomous vehicles for parking violations. The half-trillion-dollar data-center bet on a next-word predictor becoming a conscious being is unequivocally stupid. The other two are the consequences of a political economy that has decided that the cost of extracting surplus from an overbuilt infrastructure is someone else’s problem.

The public consultation on the AI revenue lemma remains open. There is no deadline, because the financial system that is funding the backstop does not believe deadlines apply to it. The arithmetic, however, is indifferent. The bezzle is a loan against the future; the future always collects. The work is to be done.